13 July 2019
TradeTalks: AI/ML for Asset & Investment Management
In this episode of Nasdaq TradeTalks, Peter Vaihansky, SVP at DataArt, speaks with Jill Malandrino about artificial intelligence and machine learning. Peter discusses operational and technological challenges surrounding these technologies such as transparency of algorithms, integration into business environments, data processing, and talent shortage.
“In specific environments, regulated environments when companies rely on very complex and poorly understood algorithms that are often trained on poorly-curated, often biased datasets, that drives decision making that is not only not objective, very subjective, very opaque. You could very well be in violation of regulations and even illegal. We're talking about environments such as … consumer credit or mortgage lending or HR. So, we think that boards of directors and shareholders and executive leadership of companies that are relying on AI in those areas should be aware of those risks because they are potentially opening themselves up to regulatory scrutiny fines, financial losses or even legal action.”
“Technology speaking, a lot of the cutting-edge research and technology comes out of academic institutions. And so, it's not exactly enterprise-ready software. It lacks things like usability or security. There's a lot of retooling that needs to take place before you're ready to deploy them in enterprise environments.”
“The other thing is that for performance reasons that software often relies on specialized hardware like GPUs or FPGAs. And so, in order to achieve proper parallelism and operability and portability, you need to do a lot of engineering around that software. Operationalizing it, making sure it's well integrated with downstream systems that are able to consume the output of your ML algorithms and generate functionality down the value chain is also very important.”
“Data is another huge area of challenge. So, if you're doing AI and ML you really need to fuse data from different data sets and data sources; your transactions, your interactions with your client, your rich media or satellite images or social media data. So, you're connecting the dots across different data sets and data sources that are both internal and external to your organization. In order to do that you need a very robust and resilient and well architected data platform that is able to ingest any data from any source any format any time any volume.
Many firms are not there with that capability. Not by far. In fact, 80 to 90 percent of work in any machine learning project or initiative is data pre-processing; identifying the data, ingesting it, cleansing it, staging it, setting up data pipelines, making it available for your analysts and your data scientists to run their analytics on. And it's very tough, it's legitimately difficult work fraught with many very hard engineering problems.”
“Technologies have not been around long enough to generate a coterie of people who you can just go to the market and hire... So, this AI worker of the future is an interesting combination of software skills, data skills, and also math to understand and design models and algorithms and that's not a very common combination in the market today.”
Learn more at Trade Talk youtube-channel here.