Lifelong learning, handling surprise, learning with less labels

To get computers to learn from tasks they perform, CVT is developing lifelong learning algorithms and systems with machine learning capabilities to continuously learn over a stream of tasks (e.g., classification, autonomy tasks etc., over its lifetime) by leveraging the learned knowledge between related tasks to generalize to other tasks. To achieve the desired lifelong learning goals, CVT is working to solve the challenges of catastrophic forgetting, which is the problem of computers starting each task with no prior knowledge. 

In the DARPA LwLL (Learning with Less Labels) and others programs, SRI has developed multiple techniques for learning from very little data.  SRI has developed a Domain Adaptive Active Meta-Learning approach through active sampling of data to maximize information gain from each sample. We have also developed Hybrid Consistency Training to jointly leverage interpolation consistency, including interpolating hidden features, that imposes linear behavior locally and data augmentation consistency that learns robust embeddings against sample variations. We also use unlabeled examples to iteratively normalize features and adapt prototypes, as opposed to commonly used one-time update, for more reliable prototype-based transductive inference. Finally, we propose a modular adaptation method that selectively performs multiple state-of-the-art adaptation methods in sequence. As different downstream tasks may require different types of adaptation, our modular adaptation enables the dynamic configuration of the most suitable modules based on the downstream task. Our methods improve over the state of the art.

In the DARPA SAIL ON (Science of Artificial Intelligence and Learning for Open-world Novelty program), CVT developed predictive coding-based techniques for detection of novelties in new worlds. We have developed symbolic-based planning to detect novel situations in the environment. For the visual domain, we exploit the Open-AI CLIP neural model to generate embedding features. We use these embedding features to detect out distribution entities and visual novelties.  

Robust AI 

Deep learning networks are vulnerable to so-called Trojan attacks in which the adversary poisons the training data with specially designed patterns that induce erratic behavior in the network when certain data, termed “triggers, “ are presented to it. CVT developed reverse engineering techniques to reconstruct possible trojan triggers, and studied the behavior of networks in response to such triggers so as to detect whether they have been the victims of a trojan attack. SRI’s techniques have been applied to Image Classification and Question answering and sentiment extraction using text. SRI’s techniques exploit logical constraints that impose compactness on the networks as well as sparsity of the attacks since they aim to go undetected. SRI’s techniques have advanced the state of the art in trojan detection and have achieved results that rank in the top three performers consistently. 

On DARPA GARD (Guaranteeing AI Robustness Against Deception), CVT addressed the need for developing methods for building attack-agnostic robustness into machine learning (ML) architectures.  Using an information-theoretic analysis framework, SRI’s GARD program developed multiple worst-case theoretical guarantees underlying different defensive approaches. Our metrics enable ranking of state-of-the-art (SOTA) Deep Neural Networks (DNNs) in terms of robustness and have paved the way towards safety standards for DNNs deployed in critical DoD applications.  

• Researchers have demonstrated effective attacks on ML algorithms.  On DARPA QED (Quantifying Ensemble Diversity for Robust Machine Learning), CVT researched developing effective defenses against such attacks which is essential if ML is to be used for defense, security, or health and safety applications.  CVT developed an approach to dynamically change network DDNN eights and activations based on stochastic quantization (or generator) of a pre-trained neural network.  We showed our moving ensemble approach improved protection of a network against attacks.