A5036343150- Neural dynamics and brain function
- Memory and Neural Mechanisms
- Neural and Behavioral Psychology Studies
- Visual perception and processing mechanisms
- Topic Modeling
- Neuroscience and Neural Engineering
- Domain Adaptation and Few-Shot Learning
- Biomedical Research and Pathophysiology
- Lysosomal Storage Disorders Research
- Advanced Neural Network Applications
- EEG and Brain-Computer Interfaces
- Cell Image Analysis Techniques
- Natural Language Processing Techniques
- Explainable Artificial Intelligence (XAI)
- Anomaly Detection Techniques and Applications
- Adversarial Robustness in Machine Learning
- Smart Grid Security and Resilience
- Neural Networks and Applications
- Calcium signaling and nucleotide metabolism
- Corruption and Economic Development
- Machine Learning in Materials Science
- Genetic and Kidney Cyst Diseases
- Viral gastroenteritis research and epidemiology
- Machine Learning and Data Classification
- Reservoir Engineering and Simulation Methods
Amgen (United States)
2025
Mosaic Event Management (United States)
2022-2024
McGill University
2011-2023
Meta (Israel)
2020-2021
Western University
2017-2019
Robarts Clinical Trials
2019
Synageva BioPharma (United States)
2011
Abstract Convolutional architectures have proven to be extremely successful for vision tasks. Their hard inductive biases enable sample-efficient learning, but come at the cost of a potentially lower performance ceiling. Vision transformers rely on more flexible self-attention layers, and recently outperformed CNNs image classification. However, they require costly pre-training large external datasets or distillation from pre-trained convolutional networks. In this paper, we ask following...
Neurons in the primate lateral prefrontal cortex (LPFC) encode working memory (WM) representations via sustained firing, a phenomenon hypothesized to arise from recurrent dynamics within ensembles of interconnected neurons. Here, we tested this hypothesis by using microelectrode arrays examine spike count correlations (rsc ) LPFC neuronal during spatial WM task. We found pattern pairwise rsc maintenance indicative stronger coupling between similarly tuned neurons and increased inhibition...
Working memory (WM) is the ability to maintain and manipulate information 'in mind'. The neural codes underlying WM have been a matter of debate. We simultaneously recorded activity hundreds neurons in lateral prefrontal cortex male macaque monkeys during visuospatial task that required navigation virtual 3D environment. Here, we demonstrate distinct neuronal activation sequences (NASs) encode remembered target locations This NAS code outperformed persistent firing for reality task, but not...
Covalent inhibition of the KRASG12C oncoprotein has emerged as a promising therapeutic approach for treatment nonsmall cell lung cancer (NSCLC). The identification inhibitors typically relied on high-throughput screening (HTS) libraries cysteine-reactive small molecules or attachment warheads to noncovalent binders KRAS. Such approaches have historically been limited in size and diversity that could be effectively screened. DNA-encoded library (DEL) accelerate preparation incredibly large...
Abstract Activating mutations in KRAS occur greater than twenty percent of all cancers encompassing several solid tumors with significant unmet medical need, including lung, colon, pancreatic, endometrial and stomach cancers. While codon 12 (encoding glycine) are the most common, activating missense also observed at codons 13 61. Guided by our experience successful development covalent G12C mutant-selective inhibitor, sotorasib, we leveraged structure-based design to identify a new class...
Single neurons in primate dorsolateral prefrontal cortex (dLPFC) are known to encode working memory (WM) representations of visual space. Psychophysical studies have shown that the horizontal and vertical meridians field can bias spatial information maintained WM. However, most models tacitly assumed dLPFC represent mnemonic space homogenously. The anatomical organization these has also eluded clear parametric description. We investigated issues by recording from neuronal ensembles macaque...
Methods for understanding the decisions of and mechanisms underlying deep neural networks (DNNs) typically rely on building intuition by emphasizing sensory or semantic features individual examples. For instance, methods aim to visualize components an input which are "important" a network's decision, measure properties single neurons. Here, we argue that interpretability research suffers from over-reliance intuition-based approaches risk-and in some cases have caused-illusory progress...
Neurons within the primate dorsolateral prefrontal cortex (dlPFC) are clustered in microcolumns according to their visuospatial tuning. One issue that remains poorly investigated is how this anatomical arrangement influences functional interactions between neurons during behavior. To investigate question we implanted 4 mm×4 mm multielectrode arrays two macaques' dlPFC area 8a and measured spike count correlations (rsc) responses of simultaneously recorded when animals maintained stationary...
The properties of individual neurons are often analyzed in order to understand the biological and artificial neural networks which they're embedded. Class selectivity-typically defined as how different a neuron's responses across classes stimuli or data samples-is commonly used for this purpose. However, it remains an open question whether is necessary and/or sufficient deep (DNNs) learn class selectivity units. We investigated causal impact on network function by directly regularizing...
Neurons in the lateral prefrontal cortex (LPFC) encode sensory and cognitive signals, as well commands for goal-directed actions. Therefore, LPFC might be a good signal source goal-selection brain-computer interface (BCI) that decodes intended goal of motor action previous to its execution. As first step development BCI, we set out determine if could decode simple behavioral intentions direct gaze eight different locations space from single-trial neural activity. We recorded neuronal spiking...
Abstract Lateral prefrontal cortex (LPFC) neurons signal the allocation of voluntary attention; however, neural computations underlying this function remain unknown. To investigate this, we recorded from neuronal ensembles in LPFC two Macaca fascicularis performing a visuospatial attention task. responses to single stimulus were normalized when additional stimuli/distracters appeared across visual field and well-characterized by an averaging computation. Deploying toward individual...
Most interpretability research in NLP focuses on understanding the behavior and features of a fully trained model. However, certain insights into model may only be accessible by observing trajectory training process. We present case study syntax acquisition masked language models (MLMs) that demonstrates how analyzing evolution interpretable artifacts throughout deepens our emergent behavior. In particular, we Syntactic Attention Structure (SAS), naturally emerging property MLMs wherein...
In this work, we investigate whether small language models can determine high-quality subsets of large-scale text datasets that improve the performance larger models. While existing work has shown pruning based on perplexity a model yield data, smaller be used for perplexity-based and how is affected by domain composition data being pruned. We demonstrate multiple dataset compositions, pretraining \emph{significantly} downstream task performance: perplexities computed with 125 million...
While the relative trade-offs between sparse and distributed representations in deep neural networks (DNNs) are well-studied, less is known about how these apply to of semantically-meaningful information. Class selectivity, variability a unit's responses across data classes or dimensions, one way quantifying sparsity semantic representations. Given recent evidence showing that class selectivity can impair generalization, we sought investigate whether it also confers robustness (or...
Working memory is the ability to briefly remember and manipulate information after it becomes unavailable senses. The mechanisms supporting working coding in primate brain remain controversial. Here we demonstrate that microcircuits layers 2/3 of lateral prefrontal cortex dynamically represent content a naturalistic task through sequential activation single neurons. We simultaneously recorded activity hundreds neurons macaque monkeys during visuospatial set virtual environment. found encoded...
Lateral prefrontal cortex (lPFC) neurons are thought to encode working memory (WM) representations of visual space via sustained firing. Neurophysiological studies WM typically record from individual neurons, thus we lack an understanding how larger ensembles simultaneously-recorded represent WM: do not know if representation fidelity is affected by phenomena that measurable in single (e.g. noise correlations– rnoise), nor coding properties scale with the size and composition neuronal...
The allocation of attention can be decoded from the activity lateral prefrontal cortex neuronal ensembles (Tremblay et al., 2015). One issue that remains unclear is impact a neural population's size and composition on decoding attention. To investigate this, we recorded responses neurons in two macaques using microelectrode arrays while they performed visuospatial task. During task, animals had to direct cued target stimulus positioned one four visual quadrants ignoring 3 identical...
It has been shown that neurons in the primate prefrontal cortex encode information during delay period of working memory tasks for spatial locations, however it remains unclear whether and how these units interact with each other. We examined this question by recording single cell responses from microelectrode arrays (each array = 96 microelectrodes, Blackrock Inc., UT) implanted dorsolateral (area 8v) two macaca fascicularis a task, measuring spike count correlations between recorded units....
In many practical scenarios -- like hyperparameter search or continual retraining with new data related training runs are performed times in sequence. Current practice is to train each of these models independently from scratch. We study the problem exploiting computation invested previous reduce cost future using knowledge distillation (KD). find that augmenting KD dramatically reduces time necessary models, even taking into account overhead KD. improve on results two strategies by 80-90%...