International Business Machines Corporation

Rensselaer Polytechnic Institute

Massachusetts Institute of Technology

Battelle Memorial Institute

University of Illinois

National Tsing Hua Universitiy

National Chung Hsing University

Pin-Yu Chen

Chuang Gan

Sijia Liu

Chia-Yu Chen

Payel Das

Bo Wu

Yada Zhu

Lingfei Wu

Amit Dhurandhar

Karthikeyan Shanmugam

Gaoyuan Zhang

Jinjun Xiong

Enara C Vijil

Brian Leo Quanz

Kumar Bhaskaran

Ronny Luss

Zhenfang Chen

Tejaswini Pedapati

Prasanna Sattigeri

Akhilan Boopathy

Songtao Lu

Ruchir Puri

Xuan Liu

Ajay Ashok Deshpande

Dakuo Wang

Xin Zhang

Yang Zhang

I-Hsin Chung

Jae-Wook Ahn

Karthikeyan Natesan Ramamurthy

Mingu Kang

Rameswar Panda

Jintao Zhang

Aurelie C Lozano

Chia-Yi Hsu

Kun Xu

Georgios Kollias

Pingkun Yan

Luca Daniel

Indrajit Ray

Igor Melnyk

Kahini Wadhawan

Samuel Chung Hoffman

Khushbu Agarwal

Avinash Balakrishnan

Bo Li

Subhajit Chaudhury

Sutanay Choudhury

Pu Zhao

Ren Wang

Tsui-Wei Weng

Meng Wang

Maurício Gruppi

Sibel Adali

Yunan Ye

Chunheng Jiang

Jianxi Gao

Jiajin Zhang

Hanqing Chao

Yue Cao

Tsung-Yi Ho

Yun-Yun Tsai

Elliot Nelson

Chia-Mu Yu

A trained machine learning model and a training dataset used to train the trained machine learning model can be received. Based on the training dataset, unsupervised adversarial examples can be generated. Robustness of the trained machine learning model can be determined using the generated unsupervised adversarial examples. The training dataset can be augmented with the generated unsupervised adversarial examples. The trained machine learning model can be retrained using the augmented training dataset.

Generating unsupervised adversarial examples for machine learning

A machine learning model is trained using original source domain data through empirical risk minimization and a model sensitivity map is computed. Each sensitive frequency point on the model sensitivity map is targeted. An adversarial technique is employed to generate spectral adversarial images based on the model sensitivity map and an image amplitude spectrum is augmented. The generated spectral adversarial images are mixed with the original source domain data to finetune the machine learning model and deployment of the finetuned machine learning model is facilitated.

Spectral data augmentation for single domain generalization

Mechanisms are provided for performing machine learning (ML) training of a ML action recognition computer model which involves processing an original input dataset to generate an object feature bank comprising object feature data structures for a plurality of different objects. For an input video, a verb data structure and an original object data structure are generated and a candidate object feature data structure is selected from the object feature bank for generation of pseudo composition (PC) training data. The PC training data is generated based on the selected candidate object feature data structure and comprises a combination of the verb data structure and the candidate object feature data structure. The PC training data represents a combination of an action and an object not represented in the original input dataset. ML training of the ML action recognition computer model is performed based on an unseen combination comprising the PC training data.

Compositional action machine learning mechanisms

Scalable distributed adversarial training techniques for robust deep neural networks are provided. In one aspect, a method for adversarial training of a deep neural network-based model by distributed computing machines M includes, by distributed computing machines M: obtaining adversarial perturbation-modified training examples for samples in a local dataset D; computing gradients of a local cost function ƒwith respect to parameters θ of the deep neural network-based model using the adversarial perturbation-modified training examples; transmitting the gradients of the local cost function ƒto a server which aggregates the gradients of the local cost function ƒand transmits an aggregated gradient to the distributed computing machines M; and updating the parameters θ of the deep neural network-based model stored at each of the distributed computing machines M based on the aggregated gradient received from the server. A method for distributed adversarial training of a deep neural network-based model by the server is also provided.

Distributed adversarial training for robust deep neural networks

A computerized system includes a generative large language model (LLM). The LLM includes an encoder configured to provide a vector input to a content addressable memory system and a decoder configured to receive a vector output from the content addressable memory system and generate a data output. The content addressable memory system is configured to receive the vector input and generate a vector output based on data contents of the content addressable memory system. The content addressable memory system is configured to respond to receiving at least one memory update command by updating the data contents of the content addressable memory system.

Dynamic updating of content addressable associative memories for large language models

An output layer is removed from a pre-trained neural network model and a neural capacitance probe unit with multiple layers is incorporated on top of one or more bottom layers of the pre-trained neural network model. The neural capacitance probe unit is randomly initialized and a modified neural network model is trained by fine-tuning the one or more bottom layers on a target dataset for a maximum number of epochs, the modified neural network model comprising the neural capacitance probe unit incorporated with multiple layers on top of the one or more bottom layers of the pre-trained neural network model. An adjacency matrix is obtained from the initialized neural capacitance probe unit and a neural capacitance metric is computed using the adjacency matrix. An active model is selected using the neural capacitance metric and a machine learning system is configured using the active model.

Neural capacitance: neural network selection via edge dynamics

A system and method for designing protein sequences conditioned on a specific target fold. The system is a transformer-based generative framework for modeling a complex sequence-structure relationship. To mitigate the heterogeneity between the sequence domain and the fold domain, a Fold-to-Sequence model jointly learns a sequence embedding using a transformer and a fold embedding from the density of secondary structural elements in 3D voxels. The joint sequence-fold representation through novel intra-domain and cross-domain losses with an intra-domain loss forces two semantically similar (where the proteins should have the same fold(s)) samples from the same domain to be close to each other in a latent space, while a cross-domain loss forces two semantically similar samples in different domains to be closer. In an embodiment, the Fold-to-Sequence model performs design tasks that include low resolution structures, structures with a region of missing residues, and NMR structural ensembles.

Embedding-based generative model for protein design

A query-based generic end-to-end molecular optimization ('QMO') system framework, method and computer program product for optimizing molecules, such as for accelerating drug discovery. The QMO framework decouples representation learning and guided search and applies to any plug-in encoder-decoder with continuous latent representations. QMO framework directly incorporates evaluations based on chemical modeling, analysis packages, and pre-trained machine-learned prediction models for efficient molecule optimization using a query-based guided search method based on zeroth order optimization. The QMO features efficient guided search with molecular property evaluations and constraints obtained using the predictive models and chemical modeling and analysis packages. QMO tasks include optimizing drug-likeness and penalized log P scores with similarity constraints and improving the target binding affinity of existing drugs to pathogens such as the SARS-CoV-2 main protease protein while preserving the desired drug properties. QMO tasks further improves optimizing antimicrobial peptides toward lower toxicity.

Query-based molecule optimization and applications to functional molecule discovery

A system may include a memory and a processor in communication with the memory. The processor may be configured to perform operations. The operations may include receiving an input, extracting features from the input, and mining object relations using the features. The operations may include determining feature vectors using the object relations and generating, using the feature vectors, an output indicating a target region, wherein the target region corresponds to the input.

Image grounding with modularized graph attentive networks

An adversarial robustness testing method, system, and computer program product include testing, via an accelerator, a robustness of a black-box system under different access settings, where the testing includes tearing down the robustness testing to a subtask of a predetermined size.

Testing adversarial robustness of systems with limited access

Transfer learning in machine learning can include receiving a machine learning model. Target domain training data for reprogramming the machine learning model using transfer learning can be received. The target domain training data can be transformed by performing a transformation function on the target domain training data. Output labels of the machine learning model can be mapped to target labels associated with the target domain training data. The transformation function can be trained by optimizing a parameter of the transformation function. The machine learning model can be reprogrammed based on input data transformed by the transformation function and a mapping of the output labels to target labels.

Transfer learning with machine learning systems

One or more computer processors improve action recognition by removing inference introduced by visual appearances of objects within a received video segment. The one or more computer processors extract appearance information and structure information from a received video segment. The one or more computer processors calculate a factual inference (TE) for the received video segment utilizing the extracted appearance information and structure information. The one or more computer processors calculate a counterfactual debiasing inference (NDE) for the received video segment. The one or more computer processors calculate a total indirect effect (TIE) by subtracting the calculated counterfactual debiased inference from the calculated factual inference. The one or more computer processors action recognize the received video segment by selecting a classification result associated with a highest calculated TIE.

Counterfactual debiasing inference for compositional action recognition

A processor training a reinforcement learning model can include receiving a first dataset representing an observable state in reinforcement learning to train a machine to perform an action. The processor receives a second dataset. Using the second dataset, the processor trains a machine learning classifier to make a prediction about an entity related to the action. The processor extracts an embedding from the trained machine learning classifier, and augments the observable state with the embedding to create an augmented state. Based on the augmented state, the processor trains a reinforcement learning model to learn a policy for performing the action, the policy including a mapping from state space to action space.

State-augmented reinforcement learning

Techniques that facilitate graph similarity analytics are provided. In one example, a system includes an information component and a similarity component. The information component generates a first information index indicative of a first entropy measure for a first graph-structured dataset associated with a machine learning system. The information component also generates a second information index indicative of a second entropy measure for a second graph-structured dataset associated with the machine learning system. The similarity component determines similarity between the first graph-structured dataset and the second graph-structured dataset based on a graph similarity computation associated with the first information index and the second information index.

Graph similarity analytics

Generate, for each of the words of a common vocabulary of first and second text corpora, a first word embedding vector in the first text corpus and a second word embedding vector in the second text corpus. Generate, for each word in a random sample of non-landmark words, an artificially shifted word embedding vector by modifying the first word embedding vector for that word. Train a machine learning classifier to predict whether an artificial shift has been injected for a given word, based on the artificially shifted word embedding vector and the second word embedding vector for the given word. Predict semantic shifts for at least a plurality of the words of the common vocabulary by providing the first word embedding vectors and the second word embedding vectors for at least the plurality of the words of the common vocabulary as input to the trained machine learning classifier.

Self-supervised semantic shift detection and alignment

Growing community of inventors

Yorktown Heights, NY, United States of America

Pin-Yu Chen