ICLR 2024 Papers

Reinforcement learning (RL) over text representations can be effective for finding high-value policies that can search over graphs. However, RL requires careful structuring of the search space and algorithm design to be effective in this challenge. Through extensive experiments, we explore how different design choices for text grammar and algorithmic choices for training can affect an RL policy's ability to generate molecules with desired properties. We arrive at a new RL-based molecular design algorithm (ChemRLformer) and perform a thorough analysis using 25 molecule design tasks, including computationally complex protein docking simulations. From this analysis, we discover unique insights in this problem space and show that ChemRLformer achieves state-of-the-art performance while being more straightforward than prior work by demystifying which design choices are actually helpful for text-based molecule design.

A graph is a structure made up of vertices and edges used to represent complex relationships between entities, while a graph stream is a continuous flow of graph updates that convey evolving relationships between entities. The massive volume and high dynamism of graph streams promote research on data structures of graph summarization, which provides a concise and approximate view of graph streams with sub-linear space and linear construction time, enabling real-time graph analytics in various domains, such as social networking, financing, and cybersecurity. In this work, we propose the Mayfly, the first neural data structure for summarizing graph streams. The Mayfly replaces handcrafted data structures with better accuracy and adaptivity. To cater to practical applications, Mayfly incorporates two offline training phases. During the larval phase, the Mayfly learns basic summarization abilities from automatically and synthetically constituted meta-tasks, and in the metamorphosis phase, it rapidly adapts to real graph streams via meta-tasks. With specific configurations of information pathways, the Mayfly enables flexible support for miscellaneous graph queries, including edge, node, and connectivity queries. Extensive empirical studies show that the Mayfly significantly outperforms its handcrafted competitors.

Scaling Transformers to longer sequence lengths has been a major problem in the last several years, promising to improve performance in language modeling and high-resolution image understanding, as well as to unlock new applications in code, audio, and video generation. The attention layer is the main bottleneck in scaling to longer sequences, as its runtime and memory increase quadratically in the sequence length. FlashAttention [5] exploits the asymmetric GPU memory hierarchy to bring significant memory saving (linear instead of quadratic) and runtime speedup (2-4× compared to optimized baselines), with no approximation. However, FlashAttention is still not nearly as fast as optimized matrix-multiply (GEMM) operations, reaching only 25-40% of the theoretical maximum FLOPs/s. We observe that the inefficiency is due to suboptimal work partitioning between different thread blocks and warps on the GPU, causing either low-occupancy or unnecessary shared memory reads/writes. We propose FlashAttention-2, with better work partitioning to address these issues. In particular, we (1) tweak the algorithm to reduce the number of non-matmul FLOPs (2) parallelize the attention computation, even for a single head, across different thread blocks to increase occupancy, and (3) within each thread block, distribute the work between warps to reduce communication through shared memory. These yield around 2× speedup compared to FlashAttention, reaching 50-73% of the theoretical maximum FLOPs/s on A100 and getting close to the efficiency of GEMM operations. We empirically validate that when used end-to-end to train GPT-style models, FlashAttention-2 reaches training speed of up to 225 TFLOPs/s per A100 GPU (72% model FLOPs utilization).

Understanding the functional organization of higher visual cortex is a central focus in neuroscience. Past studies have primarily mapped the visual and semantic selectivity of neural populations using hand-selected stimuli, which may potentially bias results towards pre-existing hypotheses of visual cortex functionality. Moving beyond conventional approaches, we introduce a data-driven method that generates natural language descriptions for images predicted to maximally activate individual voxels of interest. Our method -- Semantic Captioning Using Brain Alignments ("BrainSCUBA") -- builds upon the rich embedding space learned by a contrastive vision-language model and utilizes a pre-trained large language model to generate interpretable captions. We validate our method through fine-grained voxel-level captioning across higher-order visual regions. We further perform text-conditioned image synthesis with the captions, and show that our images are semantically coherent and yield high predicted activations. Finally, to demonstrate how our method enables scientific discovery, we perform exploratory investigations on the distribution of "person" representations in the brain, and discover fine-grained semantic selectivity in body-selective areas. Unlike earlier studies that decode text, our method derives *voxel-wise captions of semantic selectivity*. Our results show that BrainSCUBA is a promising means for understanding functional preferences in the brain, and provides motivation for further hypothesis-driven investigation of visual cortex.

The study of decoding visual neural information faces challenges in generalizing single-subject decoding models to multiple subjects, due to individual differences. Moreover, the limited availability of data from a single subject has a constraining impact on model performance. Although prior multi-subject decoding methods have made significant progress, they still suffer from several limitations, including difficulty in extracting global neural response features, linear scaling of model parameters with the number of subjects, and inadequate characterization of the relationship between neural responses of different subjects to various stimuli. To overcome these limitations, we propose a CLIP-guided Multi-sUbject visual neural information SEmantic Decoding (CLIP-MUSED) method. Our method consists of a Transformer-based feature extractor to effectively model global neural representations. It also incorporates learnable subject-specific tokens that facilitates the aggregation of multi-subject data without a linear increase of parameters. Additionally, we employ representational similarity analysis (RSA) to guide token representation learning based on the topological relationship of visual stimuli in the representation space of CLIP, enabling full characterization of the relationship between neural responses of different subjects under different stimuli. Finally, token representations are used for multi-subject semantic decoding. Our proposed method outperforms single-subject decoding methods and achieves state-of-the-art performance among the existing multi-subject methods on two fMRI datasets. Visualization results provide insights into the effectiveness of our proposed method. Code is available at https://github.com/CLIP-MUSED/CLIP-MUSED.

In this work we describe successor heads: attention heads that increment tokens with a natural ordering, such as numbers, months, and days. For example, successor heads increment 'Monday' into 'Tuesday'. We explain the successor head behavior with an approach rooted in mechanistic interpretability, the field that aims to explain how models complete tasks in human-understandable terms. Existing research in this area has struggled to find recurring, mechanistically interpretable large language model (LLM) components beyond small toy models. Further, existing results have led to very little insight to explain the internals of the larger models that are used in practice. In this paper, we analyze the behavior of successor heads in LLMs and find that they implement abstract representations that are common to different architectures. Successor heads form in LLMs with as few as 31 million parameters, and at least as many as 12 billion parameters, such as GPT-2, Pythia, and Llama-2. We find a set of 'mod 10' features that underlie how successor heads increment in LLMs across different architectures and sizes. We perform vector arithmetic with these features to edit head behavior and provide insights into numeric representations within LLMs. Additionally, we study the behavior of successor heads on natural language data, where we find that successor heads are important for achieving a low loss on examples involving succession, and also identify interpretable polysemanticity in a Pythia successor head.

Measuring geometric similarity between high-dimensional network representations is a topic of longstanding interest to neuroscience and deep learning. Although many methods have been proposed, only a few works have rigorously analyzed their statistical efficiency or quantified estimator uncertainty in data-limited regimes. Here, we derive upper and lower bounds on the worst-case convergence of standard estimators of shape distance—a measure of representational dissimilarity proposed by Williams et al. (2021). These bounds reveal the challenging nature of the problem in high-dimensional feature spaces. To overcome these challenges, we introduce a novel method-of-moments estimator with a tunable bias-variance tradeoff parameterized by an upper bound on bias. We show that this estimator achieves superior performance to standard estimators in simulation and on neural data, particularly in high-dimensional settings. Our theoretical work and estimator thus respectively define and dramatically expand the scope of neural data for which geometric similarity can be accurately measured.

Reliably reconstructing physical fields from sparse sensor data is a challenge that frequenty arises in many scientific domains. In practice, the process generating the data is often not known to sufficient accuracy. Therefore, there is a growing interest in the deep neural network route to the problem. In this work, we present a novel approach that learns a continuous representation of the field using implicit neural representations (INR). Specifically, after factorizing spatiotemporal variability into spatial and temporal components using the technique of separation of variables, the method learns relevant basis functions from sparsely sampled irregular data points to thus develop a continuous representation of the data. In experimental evaluations, the proposed model outperforms recent INR methods, offering superior reconstruction quality on simulation data from a state of the art climate model and on a second dataset that comprises of ultra-high resolution satellite-based sea surface temperature field. [Website for the Project: Both data and code are accessible.](https://xihaier.github.io/ICLR-2024-MMGN/)

Spiking Neural Networks (SNNs) are attracting growing interest for their energy-efficient computing when implemented on neuromorphic hardware. However, directly training SNNs, even adopting batch normalization (BN), is highly challenging due to their non-differentiable activation function and the temporally delayed accumulation of outputs over time. For SNN training, this temporal accumulation gives rise to Temporal Covariate Shifts (TCS) along the temporal dimension, a phenomenon that would become increasingly pronounced with layer-wise computations across multiple layers and multiple time-steps. In this paper, we introduce TAB (Temporal Accumulated Batch Normalization), a novel SNN batch normalization method that addresses the temporal covariate shift issue by aligning with neuron dynamics (specifically the accumulated membrane potential) and utilizing temporal accumulated statistics for data normalization. Within its framework, TAB effectively encapsulates the historical temporal dependencies that underlie the membrane potential accumulation process, thereby establishing a natural connection between neuron dynamics and TAB batch normalization. Experimental results on CIFAR-10, CIFAR-100, and DVS-CIFAR10 show that our TAB method outperforms other state-of-the-art methods.

Protein-Protein Interactions (PPIs) are fundamental in various biological processes and play a key role in life activities. The growing demand and cost of experimental PPI assays require computational methods for efficient PPI prediction. While existing methods rely heavily on protein sequence for PPI prediction, it is the protein structure that is the key to determine the interactions. To take both protein modalities into account, we define the microenvironment of an amino acid residue by its sequence and structural contexts, which describe the surrounding chemical properties and geometric features. In addition, microenvironments defined in previous work are largely based on experimentally assayed physicochemical properties, for which the "vocabulary" is usually extremely small. This makes it difficult to cover the diversity and complexity of microenvironments. In this paper, we propose Microenvironment-Aware Protein Embedding for PPI prediction (MPAE-PPI), which encodes microenvironments into chemically meaningful discrete codes via a sufficiently large microenvironment "vocabulary" (i.e., codebook). Moreover, we propose a novel pre-training strategy, namely Masked Codebook Modeling (MCM), to capture the dependencies between different microenvironments by randomly masking the codebook and reconstructing the input. With the learned microenvironment codebook, we can reuse it as an off-the-shelf tool to efficiently and effectively encode proteins of different sizes and functions for large-scale PPI prediction. Extensive experiments show that MAPE-PPI can scale to PPI prediction with millions of PPIs with superior trade-offs between effectiveness and computational efficiency than the state-of-the-art competitors.

Sparse training (ST) aims to ameliorate deep learning by replacing fully connected artificial neural networks (ANNs) with sparse or ultra-sparse ones, such as brain networks are, therefore it might benefit to borrow brain-inspired learning paradigms from complex network intelligence theory. Here, we launch the ultra-sparse advantage challenge, whose goal is to offer evidence on the extent to which ultra-sparse (around 1\% connection retained) topologies can achieve any leaning advantage against fully connected. Epitopological learning is a field of network science and complex network intelligence that studies how to implement learning on complex networks by changing the shape of their connectivity structure (epitopological plasticity). One way to implement Epitopological (epi- means new) Learning is via link prediction: predicting the likelihood of non-observed links to appear in the network. Cannistraci-Hebb learning theory inspired the CH3-L3 network automata rule for link prediction which is effective for general-purpose link prediction. Here, starting from CH3-L3 we propose Epitopological Sparse Meta-deep Learning (ESML) to apply Epitopological Learning to sparse training. In empirical experiments, we find that ESML learns ANNs with ultra-sparse hyperbolic (epi-)topology in which emerges a community layer organization that is meta-deep (meaning that each layer also has an internal depth due to power-law node hierarchy). Furthermore, we discover that ESML can in many cases automatically sparse the neurons during training (arriving even to 30\% neurons left in hidden layers), this process of node dynamic removal is called percolation. Starting from this network science evidence, we design Cannistraci-Hebb training (CHT), a 4-step training methodology that puts ESML at its heart. We conduct experiments on 7 datasets and 5 network structures comparing CHT to dynamic sparse training SOTA algorithms and the fully connected counterparts. The results indicate that, with a mere 1\% of links retained during training, CHT surpasses fully connected networks on VGG16, GoogLeNet, ResNet50, and ResNet152. This key finding is an evidence for ultra-sparse advantage and signs a milestone in deep learning. CHT acts akin to a gradient-free oracle that adopts CH3-L3-based epitopological learning to guide the placement of new links in the ultra-sparse network topology to facilitate sparse-weight gradient learning, and this in turn reduces the convergence time of ultra-sparse training. Finally, CHT offers the first examples of parsimony dynamic sparse training because, in many datasets, it can retain network performance by percolating and significantly reducing the node network size. Our code is available at: https://github.com/biomedical-cybernetics/Cannistraci-Hebb-training

We extend JAX with the capability to automatically differentiate higher-order functions (functionals and operators). By representing functions as infinite dimensional generalization of arrays, we seamlessly use JAX's existing primitive system to implement higher-order functions. We present a set of primitive operators that serve as foundational building blocks for constructing several key types of functionals. For every introduced primitive operator, we derive and implement both linearization and transposition rules, aligning with JAX's internal protocols for forward and reverse mode automatic differentiation. This enhancement allows for functional differentiation in the same syntax traditionally use for functions. The resulting functional gradients are themselves functions ready to be invoked in python. We showcase this tool's efficacy and simplicity through applications where functional derivatives are indispensable.

Optimizing and certifying the positivity of polynomials are fundamental primitives across mathematics and engineering applications, from dynamical systems to operations research. However, solving these problems in practice requires large semidefinite programs, with poor scaling in dimension and degree. In this work, we demonstrate for the first time that neural networks can effectively solve such problems in a data-driven fashion, achieving tenfold speedups while retaining high accuracy. Moreover, we observe that these polynomial learning problems are equivariant to the non-compact group $SL(2,\mathbb{R})$, which consists of area-preserving linear transformations. We therefore adapt our learning pipelines to accommodate this structure, including data augmentation, a new $SL(2,\mathbb{R})$-equivariant architecture, and an architecture equivariant with respect to its maximal compact subgroup, $SO(2, \mathbb{R})$. Surprisingly, the most successful approaches in practice do not enforce equivariance to the entire group, which we prove arises from an unusual lack of architecture universality for $SL(2,\mathbb{R})$ in particular. A consequence of this result, which is of independent interest, is that there exists an equivariant function for which there is no sequence of equivariant polynomials multiplied by arbitrary invariants that approximates the original function. This is a rare example of a symmetric problem where data augmentation outperforms a fully equivariant architecture, and provides interesting lessons in both theory and practice for other problems with non-compact symmetries.

Mesh deformation plays a pivotal role in many 3D vision tasks including dynamic simulations, rendering, and reconstruction. However, defining an efficient discrepancy between predicted and target meshes remains an open problem. A prevalent approach in current deep learning is the set-based approach which measures the discrepancy between two surfaces by comparing two randomly sampled point-clouds from the two meshes with Chamfer pseudo-distance. Nevertheless, the set-based approach still has limitations such as lacking a theoretical guarantee for choosing the number of points in sampled point-clouds, and the pseudo-metricity and the quadratic complexity of the Chamfer divergence. To address these issues, we propose a novel metric for learning mesh deformation. The metric is defined by sliced Wasserstein distance on meshes represented as probability measures that generalize the set-based approach. By leveraging probability measure space, we gain flexibility in encoding meshes using diverse forms of probability measures, such as continuous, empirical, and discrete measures via \textit{varifold} representation. After having encoded probability measures, we can compare meshes by using the sliced Wasserstein distance which is an effective optimal transport distance with linear computational complexity and can provide a fast statistical rate for approximating the surface of meshes. To the end, we employ a neural ordinary differential equation (ODE) to deform the input surface into the target shape by modeling the trajectories of the points on the surface. Our experiments on cortical surface reconstruction demonstrate that our approach surpasses other competing methods in multiple datasets and metrics.

Spiking Neural Networks (SNNs) have garnered considerable attention due to their energy efficiency and unique biological characteristics. However, the widely adopted Leaky Integrate-and-Fire (LIF) model, as the mainstream neuron model in current SNN research, has been revealed to exhibit significant deficiencies in deep-layer gradient calculation and capturing global information on the time dimension. In this paper, we propose the Learnable Multi-hierarchical (LM-H) model to address these issues by dynamically regulating its membrane-related factors. We point out that the LM-H model fully encompasses the information representation range of the LIF model while offering the flexibility to adjust the extraction ratio between historical and current information. Additionally, we theoretically demonstrate the effectiveness of the LM-H model and the functionality of its internal parameters, and propose a progressive training algorithm tailored specifically for the LM-H model. Furthermore, we devise an efficient training framework for our novel advanced model, encompassing hybrid training and time-slicing online training. Through extensive experiments on various datasets, we validate the remarkable superiority of our model and training algorithm compared to previous state-of-the-art approaches. Code is available at [https://github.com/hzc1208/STBP_LMH](https://github.com/hzc1208/STBP_LMH).

Predicting physical properties of materials from their crystal structures is a fundamental problem in materials science. In peripheral areas such as the prediction of molecular properties, fully connected attention networks have been shown to be successful. However, unlike these finite atom arrangements, crystal structures are infinitely repeating, periodic arrangements of atoms, whose fully connected attention results in *infinitely connected attention*. In this work, we show that this infinitely connected attention can lead to a computationally tractable formulation, interpreted as *neural potential summation*, that performs infinite interatomic potential summations in a deeply learned feature space. We then propose a simple yet effective Transformer-based encoder architecture for crystal structures called *Crystalformer*. Compared to an existing Transformer-based model, the proposed model requires only 29.4% of the number of parameters, with minimal modifications to the original Transformer architecture. Despite the architectural simplicity, the proposed method outperforms state-of-the-art methods for various property regression tasks on the Materials Project and JARVIS-DFT datasets.

Spiking Neural Networks (SNNs) have emerged as energy-efficient alternatives to Artificial Neural Networks (ANNs) when deployed on neuromorphic chips. While recent studies have demonstrated the impressive performance of deep SNNs on challenging tasks, their energy efficiency advantage has been diminished. Existing methods targeting energy consumption reduction do not fully exploit sparsity, whereas powerful pruning methods can achieve high sparsity but are not directly targeted at energy efficiency, limiting their effectiveness in energy saving. Furthermore, none of these works fully exploit the sparsity of neurons or the potential for unstructured neuron pruning in SNNs. In this paper, we propose a novel pruning framework that combines unstructured weight pruning with unstructured neuron pruning to maximize the utilization of the sparsity of neuromorphic computing, thereby enhancing energy efficiency. To the best of our knowledge, this is the first application of unstructured neuron pruning to deep SNNs. Experimental results demonstrate that our method achieves impressive energy efficiency gains. The sparse network pruned by our method with only 0.63\% remaining connections can achieve a remarkable 91 times increase in energy efficiency compared to the original dense network, requiring only 8.5M SOPs for inference, with merely 2.19\% accuracy loss on the CIFAR-10 dataset. Our work suggests that deep and dense SNNs exhibit high redundancy in energy consumption, highlighting the potential for targeted SNN sparsification to save energy.

Large language models are powerful systems that excel at many tasks, ranging from translation to mathematical reasoning. Yet, at the same time, these models often show unhuman-like characteristics. In the present paper, we address this gap and ask whether large language models can be turned into cognitive models. We find that -- after finetuning them on data from psychological experiments -- these models offer accurate representations of human behavior, even outperforming traditional cognitive models in two decision-making domains. In addition, we show that their representations contain the information necessary to model behavior on the level of individual subjects. Finally, we demonstrate that finetuning on multiple tasks enables large language models to predict human behavior in a previously unseen task. Taken together, these results suggest that large, pre-trained models can be adapted to become models of human cognition, which opens up future research directions toward building more general cognitive models.

Electroencephalography (EEG) signals, known for convenient non-invasive acquisition but low signal-to-noise ratio, have recently gained substantial attention due to the potential to decode natural images. This paper presents a self-supervised framework to demonstrate the feasibility of learning image representations from EEG signals, particularly for object recognition. The framework utilizes image and EEG encoders to extract features from paired image stimuli and EEG responses. Contrastive learning aligns these two modalities by constraining their similarity. Our approach achieves state-of-the-art results on a comprehensive EEG-image dataset, with a top-1 accuracy of 15.6% and a top-5 accuracy of 42.8% in 200-way zero-shot tasks. Moreover, we perform extensive experiments to explore the biological plausibility by resolving the temporal, spatial, spectral, and semantic aspects of EEG signals. Besides, we introduce attention modules to capture spatial correlations, providing implicit evidence of the brain activity perceived from EEG data. These findings yield valuable insights for neural decoding and brain-computer interfaces in real-world scenarios. Code available at https://github.com/eeyhsong/NICE-EEG.