Neural networks with deep architectures have demonstrated significant\nperformance improvements in computer vision, speech recognition, and natural\nlanguage processing. The challenges in information retrieval (IR), however, are\ndifferent from these other application areas. A common form of IR involves\nranking of documents--or short passages--in response to keyword-based queries.\nEffective IR systems must deal with query-document vocabulary mismatch problem,\nby modeling relationships between different query and document terms and how\nthey indicate relevance. Models should also consider lexical matches when the\nquery contains rare terms--such as a person's name or a product model\nnumber--not seen during training, and to avoid retrieving semantically related\nbut irrelevant results. In many real-life IR tasks, the retrieval involves\nextremely large collections--such as the document index of a commercial Web\nsearch engine--containing billions of documents. Efficient IR methods should\ntake advantage of specialized IR data structures, such as inverted index, to\nefficiently retrieve from large collections. Given an information need, the IR\nsystem also mediates how much exposure an information artifact receives by\ndeciding whether it should be displayed, and where it should be positioned,\namong other results. Exposure-aware IR systems may optimize for additional\nobjectives, besides relevance, such as parity of exposure for retrieved items\nand content publishers. In this thesis, we present novel neural architectures\nand methods motivated by the specific needs and challenges of IR tasks.\n