The goal of my research from this summer was to further develop human-like intelligence in machine learning algorithms and accelerate their execution via hardware. The particular algorithms of interest belong to a subfield of artificial intelligence known as meta-learning, which means “learning to learn.” Humans learn new concepts with very little supervision. For example, a child can generalize the concept of “giraffe” from a few pictures in a book, but our best deep learning systems need hundreds or thousands of examples. Meta-learning algorithms help bridge the gap between human and machine intelligence by learning new skills and adapting to new environments rapidly with few training examples. Few-shot learning is a type of meta-learning where a model learns a class from few (< 10) labeled examples. Such “lifelong learning” models can continuously learn from small episodes of data containing various unseen classes. My role was to train one such few-shot learning algorithm, called a “matching network,” and execute inference for a classification task on in-memory computing hardware. My project succeeded in demonstrating few-shot learning with analog content addressable memory (ACAM) using less energy, space, and search time than alternative (GPU, TCAM) implementations with negligible compromise on inference accuracy.