Catastrophic forgetting -- the phenomenon of a neural network learning a task t1 and losing the ability to perform it after being trained on some other task t2 -- is a long-standing problem for neural networks [McCloskey and Cohen, 1989]. We present a method, Eidetic Learning, that provably solves catastrophic forgetting. A network trained with Eidetic Learning -- here, an EideticNet -- requires no rehearsal or replay. We consider successive discrete tasks and show how at inference time an EideticNet automatically routes new instances without auxiliary task information. An EideticNet bears a family resemblance to the sparsely-gated Mixture-of-Experts layer Shazeer et al. [2016] in that network capacity is partitioned across tasks and the network itself performs data-conditional routing. An EideticNet is easy to implement and train, is efficient, and has time and space complexity linear in the number of parameters. The guarantee of our method holds for normalization layers of modern neural networks during both pre-training and fine-tuning. We show with a variety of network architectures and sets of tasks that EideticNets are immune to forgetting. While the practical benefits of EideticNets are substantial, we believe they can be benefit practitioners and theorists alike. The code for training EideticNets is available at https://github.com/amazon-science/eideticnet-training.

Large language models (LLMs) often exhibit gender bias, posing challenges for their safe deployment. Existing methods to mitigate bias lack a comprehensive understanding of its mechanisms or compromise the model's core capabilities. To address these issues, we propose the CommonWords dataset, to systematically evaluate gender bias in LLMs. Our analysis reveals pervasive bias across models and identifies specific neuron circuits, including gender neurons and general neurons, responsible for this behavior. Notably, editing even a small number of general neurons can disrupt the model's overall capabilities due to hierarchical neuron interactions. Based on these insights, we propose an interpretable neuron editing method that combines logit-based and causal-based strategies to selectively target biased neurons. Experiments on five LLMs demonstrate that our method effectively reduces gender bias while preserving the model's original capabilities, outperforming existing fine-tuning and editing approaches. Our findings contribute a novel dataset, a detailed analysis of bias mechanisms, and a practical solution for mitigating gender bias in LLMs.

Weaknesses: The idea of applying Shapley values for the understanding of deep neural networks is not new. Several works, such as Lundberg et al., 2017, have already discussed the theoretical motivation for using Shapley values as an attribution method to rank the importance of the input features. Lundberg et al., 2017 also proposed approximations like KernelSHAP and DeepSHAP, which are not compared to TMAB-Shapley. Besides this line of works, the idea of using Shapley values to rank the internal neurons has been proposed by the Stier et al., 2018 (cited) and Florin Leon, 2014 (not cited) in the context of pruning. Finally, Ancona et al., 2019 (not cited) proposed an approximation technique for Shapley values tailored for deep neural networks.

We find arithmetic ability resides within a limited number of attention heads, with each head specializing in distinct operations. To delve into the reason, we introduce the Comparative Neuron Analysis (CNA) method, which identifies an internal logic chain consisting of four distinct stages from input to prediction: feature enhancing with shallow FFN neurons, feature transferring by shallow attention layers, feature predicting by arithmetic heads, and prediction enhancing among deep FFN neurons. Moreover, we identify the human-interpretable FFN neurons within both feature-enhancing and feature-predicting stages. These findings lead us to investigate the mechanism of LoRA, revealing that it enhances prediction probabilities by amplifying the coefficient scores of FFN neurons related to predictions. Finally, we apply our method in model pruning for arithmetic tasks and model editing for reducing gender bias. Code is on https://github.com/zepingyu0512/arithmetic-mechanism.

Language Models (LMs) acquire parametric knowledge from their training process, embedding it within their weights. The increasing scalability of LMs, however, poses significant challenges for understanding a model's inner workings and further for updating or correcting this embedded knowledge without the significant cost of retraining. This underscores the importance of unveiling exactly what knowledge is stored and its association with specific model components. Instance Attribution (IA) and Neuron Attribution (NA) offer insights into this training-acquired knowledge, though they have not been compared systematically. Our study introduces a novel evaluation framework to quantify and compare the knowledge revealed by IA and NA. To align the results of the methods we introduce the attribution method NA-Instances to apply NA for retrieving influential training instances, and IA-Neurons to discover important neurons of influential instances discovered by IA. We further propose a comprehensive list of faithfulness tests to evaluate the comprehensiveness and sufficiency of the explanations provided by both methods. Through extensive experiments and analysis, we demonstrate that NA generally reveals more diverse and comprehensive information regarding the LM's parametric knowledge compared to IA. Nevertheless, IA provides unique and valuable insights into the LM's parametric knowledge, which are not revealed by NA. Our findings further suggest the potential of a synergistic approach of combining the diverse findings of IA and NA for a more holistic understanding of an LM's parametric knowledge.

Fault-tolerant deep learning accelerator is the basis for highly reliable deep learning processing and critical to deploy deep learning in safety-critical applications such as avionics and robotics. Since deep learning is known to be computing- and memory-intensive, traditional fault-tolerant approaches based on redundant computing will incur substantial overhead including power consumption and chip area. To this end, we propose to characterize deep learning vulnerability difference across both neurons and bits of each neuron, and leverage the vulnerability difference to enable selective protection of the deep learning processing components from the perspective of architecture layer and circuit layer respectively. At the same time, we observe the correlation between model quantization and bit protection overhead of the underlying processing elements of deep learning accelerators, and propose to reduce the bit protection overhead by adding additional quantization constrain without compromising the model accuracy. Finally, we employ Bayesian optimization strategy to co-optimize the correlated cross-layer design parameters at algorithm layer, architecture layer, and circuit layer to minimize the hardware resource consumption while fulfilling multiple user constraints including reliability, accuracy, and performance of the deep learning processing at the same time.

Neural code intelligence models continue to be 'black boxes' to the human programmer. This opacity limits their application towards code intelligence tasks, particularly for applications like vulnerability detection where a model's reliance on spurious correlations can be safety-critical. We introduce a neuron-level approach to interpretability of neural code intelligence models which eliminates redundancy due to highly similar or task-irrelevant neurons within these networks. We evaluate the remaining important neurons using probing classifiers which are often used to ascertain whether certain properties have been encoded within the latent representations of neural intelligence models. However, probing accuracies may be artificially inflated due to repetitive and deterministic nature of tokens in code datasets. Therefore, we adapt the selectivity metric originally introduced in NLP to account for probe memorization, to formulate our source-code probing tasks. Through our neuron analysis, we find that more than 95\% of the neurons are redundant wrt. our code intelligence tasks and can be eliminated without significant loss in accuracy. We further trace individual and subsets of important neurons to specific code properties which could be used to influence model predictions. We demonstrate that it is possible to identify 'number' neurons, 'string' neurons, and higher level 'text' neurons which are responsible for specific code properties. This could potentially be used to modify neurons responsible for predictions based on incorrect signals. Additionally, the distribution and concentration of the important neurons within different source code embeddings can be used as measures of task complexity, to compare source-code embeddings and guide training choices for transfer learning over similar tasks.

The final result is quite sensitive to the starting random seed as reported in [1, 2], who observed 0.5% difference in accuracy between worst and best seed on Imagenet dataset and 1.8% difference on CIFAR-10 dataset. Thus, one might need to run an experiment several times to avoid hitting the unlucky seed. The final selected network is just the one with the best validation accuracy. We believe that discrepancy between starting seeds performance can be explained by selecting slightly different features in hidden layers in each initialization. One might ask a question, can we somehow select better features for network training? One approach is to train a bigger network and then select the most important channels via channel pruning [3, 4, 5, 6]. Training a big network, which is subsequently pruned, might be in many cases prohibitive, since increasing network width by a factor of two results in a four times increase in FLOPs and also might require a change in some hyperparameters (e.g.

DNN-based video object detection (VOD) powers autonomous driving and video surveillance industries with rising importance and promising opportunities. However, adversarial patch attack yields huge concern in live vision tasks because of its practicality, feasibility, and powerful attack effectiveness. This work proposes Themis, a software/hardware system to defend against adversarial patches for real-time robust video object detection. We observe that adversarial patches exhibit extremely localized superficial feature importance in a small region with non-robust predictions, and thus propose the adversarial region detection algorithm for adversarial effect elimination. Themis also proposes a systematic design to efficiently support the algorithm by eliminating redundant computations and memory traffics. Experimental results show that the proposed methodology can effectively recover the system from the adversarial attack with negligible hardware overhead.