||Deep Neural Networks (DNNs) are the state-of-art solution in a growing number of tasks including computer vision, speech recognition, and genomics. However, DNNs are computationally expensive as they are carefully trained to extract and abstract features from raw data using multiple layers of neurons with millions of parameters. In this dissertation, we primarily focus on inference, e.g., using a DNN to classify an input image. This is an operation that will be repeatedly performed on billions of devices in the datacenter, in self-driving cars, in drones, etc. We observe that DNNs spend a vast majority of their runtime to runtime performing matrix-by-vector multiplications (MVM). MVMs have two major bottlenecks: fetching the matrix and performing sum-of-product operations. To address these bottlenecks, we use in-situ computing, where the matrix is stored in programmable resistor arrays, called crossbars, and sum-of-product operations are performed using analog computing. In this dissertation, we propose two hardware units, ISAAC and Newton.In ISAAC, we show that in-situ computing designs can outperform DNN digital accelerators, if they leverage pipelining, smart encodings, and can distribute a computation in time and space, within crossbars, and across crossbars. In the ISAAC design, roughly half the chip area/power can be attributed to the analog-to-digital conversion (ADC), i.e., it remains the key design challenge in mixed-signal accelerators for deep networks. In spite of the ADC bottleneck, ISAAC is able to out-perform the computational efficiency of the state-of-the-art design (DaDianNao) by 8x. In Newton, we take advantage of a number of techniques to address ADC inefficiency. These techniques exploit matrix transformations, heterogeneity, and smart mapping of computation to the analog substrate. We show that Newton can increase the efficiency of in-situ computing by an additional 2x. Finally, we show that in-situ computing, unfortunately, cannot be easily adapted to handle training of deep networks, i.e., it is only suitable for inference of already-trained networks. By improving the efficiency of DNN inference with ISAAC and Newton, we move closer to low-cost deep learning that in turn will have societal impact through self-driving cars, assistive systems for the disabled, and precision medicine.