DZone Research: Keys to AI Success
DZone Research: Keys to AI Success
The keys to a successful AI strategy are to have a well-defined business problem to solve and a sound data management program to ensure you have the data to solve the business problem.
Join the DZone community and get the full member experience.Join For Free
Bias comes in a variety of forms, all of them potentially damaging to the efficacy of your ML algorithm. Read how Alegion's Chief Data Scientist discusses the source of most headlines about AI failures here.
To gather insights on the state of artificial intelligence (AI), and all of its sub-segments — machine learning (ML), natural language processing (NLP), deep learning (DL), robotic process automation (RPA), regression, et al, we talked to 21 executives who are implementing AI in their own organization and helping others understand how AI can help their business. We began by asking, "What are the keys to a successful A.I. strategy?" Here's what they told us:
- We’ve found that AI success is four-fold: 1) a clearly defined solution to a customer need; 2) cutting-edge AI/ML models; 3) subject matter experts that provide our customers advice in their particular domain; 4) a dedicated QA team to constantly evaluate the accuracy of AI models and generate training data.
- Truly understanding what you are trying to do. That kind of domain knowledge where you understand what the point is. The second task is the logistics, and then you have to do the learning too. Handle logistics, deploy models, and get feedback from them. Handling the data is a huge part of that.
- Think of your use case. Relatively consistent data based on the data. For the cost and the growth opportunity, is it able to solve a problem? Can you tap into markets that you could not have tapped into before? How much money can you make off of this? What kind of task can be automated and what type of ML is right? How much does it cost to source, test, and train the data? There is only so much you can do with data. You need the right documentation. AI/ML is a moving a chess piece. Use AI to produce reports. Unsupervised ML learns how to game the domain. 65 million videos per month and tag and identify people in the video. Lowering barrier to entry given the cost of test and training data.
- There is a massive amount of hype and companies trying to figure out what to do because they’re afraid of getting behind. They are not taking the time to define the business problem they’re trying to solve. As with anything else, the question is what the business problem is where AI will give you an unfair advantage. People are panicking because companies are figuring out how to use AI to get an unfair advantage and disrupt. It is a big element of uncertainty. A good example is Spotify since posted first report excelled at suggested playlists using AI. Only a few have picked up is that this was an acquisition that realized the strategic value and made the right acquisition. Realized weakness and acquired the right player. This is happening in a lot of different ways. Microsoft realized Slack is a better way to do communication and launched Teams as a competitor. Google released hangouts/chat tried to position Slack challenger as something that’s AI-powered.
- To receive the most value, businesses need to identify what can be leveraged from technology, what is the latest innovation delivering to them, or what can be capitalized on the innovations.
- A successful AI strategy starts with education. Executive leaders must be educated on what AI is and isn’t, what it can and cannot do, what the potential benefits and risks are to the company, the investment required — not just in economic terms, but also in terms of skills, human capital, and IT infrastructure — and the timeline to realize those benefits. Beyond education, a sound AI strategy requires a clear vision and pragmatic plan. What is the goal that you’re trying to achieve? It must tie to a specific problem or result that you’re trying to achieve. Deep domain expertise is in short supply, so the strategy should include a realistic timeline with clear and measurable outcomes tied to those key problems. Keep the problems simple. Start with small, low-risk pilot projects to get familiar with the technology and build the requisite skills and training data sets. Use the experience gained to move to more ambitious projects.
- The most adaptable brands in today’s world will thrive in the AI-enabled landscape. As explained by Harvard Business Review, the key to a successful AI strategy includes the willingness to experiment and learn quickly to develop solutions that will transform the consumer experience. A focus on user experience is certainly at the forefront as well; keeping the end user in mind is of the utmost importance in developing solutions that fit seamlessly in today’s connected world and add value. We need to view AI as a solution to problems versus a magic technology that will eliminate problems in the first place. Organizations still need many other components around their AI technology to integrate it into workflows, etc. They should also be sure to keep the problems they want to solve top of mind rather than simply saying — “We need to do something with AI.” AI and Deep Learning are best suited for tasks where you need to find patterns and useful information in data, which points to another important point: you need to have access to data in order to make AI work.
- It’s imperative that any system not be a black box. Successful and statistically rigorous systems must provide insight into the “why” behind a particular recommendation. This builds trust, which helps with adoption. As this becomes the norm, it will put pressure on all artificial intelligence solutions providers to be more rigorous and disciplined in understanding the problems they are trying to solve and the unique solutions they are building. The other aspect of successful AI deployment is first trying to seek a solution to a very fundamental problem from first principles rather than using commodity machine learning packages and tools. Machine learning is very dangerous in the hands of those who don’t understand the fundamentals behind it.
- I’ve often said these commodity packages are the high-interest credit card of technical debt. It’s very easy to get to a quick prototype, but it’s very hard to get to a stable and reliable environment without intimately understanding the problems you are solving and the technology you are using to do it.
- Clients want insights. ML is just linear regression. Enterprise customers don’t understand how to get these insights. There are not enough data scientists and no data flow pipelines. Organizations are entranced with the idea of AI but don’t know where to start and how to create, test, and deploy models. Take data and produce results. Organizations don’t know how to make the cloud work for them. We just consume fast data at the edge and use it to self-train models on the fly for every entity we create a digital twin and remember what is useful in terms of being able to predict the future. Deliver the results through real-time streaming APIs.
- We find with kind of emerging success stories with Watson making bold claims. Customers understand that AI holds some benefit, but two flavors: 1) understand the problem you want to solve and know the benefit; 2) we have a ton of data and we don’t know what the insights are, but we want to use it to improve our business. Typically, people in the first group are more successful with their AI initiative than those in the second group. For any AI/ML analytics initiative to succeed, the data house must be in order and the goals agreed upon. If no requirements are vague, we know we’re in for a longer engagement.
- It all starts with data. ML is an exercise in overcoming variation in data. The amount of data depends on the amount of variation and needs as many different variations in target and background as possible. Examples and counterexamples are also needed.
- Having a huge number of sources is critical for successful AI implementation. Because we analyze millions of minutes of live and pre-recorded audio content each day, we are able to come up with a robust industry classifier. We use ML techniques to train the platform to learn from audio cues, listen and divide audio content based on topics, and have a tone of voice, which allows us to segment and personalize audio content. It’s important to consistently train the algorithm and build classifiers so the platform can recognize the relevant keywords and offer a more accurate segmentation of audio clips.
- There are several keys to success. First, you must create a platform to collect data. Then, you need to make that data applicable to your domain. In our case, it is tying the data to wireless classifiers that measure key metrics like time to connect, throughput, roaming, and more. With the classifiers in place, you can then apply data science techniques like machine learning, mutual information, and time series anomaly detection to understand the data and convert it into meaningful insight. Finally, you want to put a face on the results. This is where a virtual assistant comes in, as it lets you ask simple language queries and it provides results in a way that is easy to understand.
- The confluence of the trends in the last few years wants to benefit from AI, but it doesn't know how. They’ve hired data scientists and bought a lot of equipment, but they’re not sure how to get from there to production. What they really need is a practice on how to put ML into production. This is not something you do only one time. Once you begin the initiative, you will be maintaining the initiative, and over time, you will add more initiative. Helping clients understand this is a new part of their operational practice. Many customers understand this instinctively, but they’re looking for someone to describe to them what the practice should look like and how to implement it. Realize approach will not scale. We’ve tried ourselves and know what we are doing will not scale in the long term.
- ML is successfully deploying AI. We came to this realization despite the hype and people wanting to do it. There are very few reach tangible business results, which is an oxymoron because so many smart people, with high salaries, don’t cross the chasm of reality. Realizing business value is not happening in line with investment and anticipation. There needs to be an operation solution for the catching side. Let data scientists do their thing, but you need an operations team knowledgeable enough to deploy and manage in an accountable manner in a 24/7 environment. We've witnessed the world shifting from frameworks to web apps. This staff function will happen with the adoption of ML. There needs to be an APM-like solution for the world of AI solutions with ML driven services.
- AI is so broad and they are many things that are part of it. We deal with enterprise AI where businesses are trying to leverage ML and DL. We see a lot of Monte Carlo and linear regression. We interface with core data-driven ML workloads. You need a holistic strategy with time spent in R&D trying to train the model. SMEs spend a lot of time training the model and less time operationalizing it. It is driven by more traditional analytics. Taking a trained model beyond the lab and operationalizing it and putting into day-to-day processes to see ROI. Less technology and more people and organizational. Silos from the business perspective hinder collaboration among business analysts, data scientists, and IT. Taking these models from data scientist and used by business analyst can drive a lot of value advantage of human-driven plus machine-driven analysis. That’s what we recommend to the client. Take down people silos to get more value.
- One of the first challenges for enterprises looking at diving into an AI-driven strategy is weighing the promised benefits of AI-enhanced productivity against initial onboarding. The opportunity costs of training employees and adjusting workflows all add up. It’s important to understand the budget holistically prior to investing in any solution, but an experienced product manager can assist with this process. In particular, staged approaches with discreet benefits for each stage will help to avoid disillusionment with AI in general. Additionally, tools are only as good as the people that use them. Some employees may be wary of technology that can affect their job, so introducing the solution as a way to augment their daily tasks is important. We believe the best way to encourage adoption by users is to clearly explain how the technology develops a solution and resolves an issue. This gives employees an “under the hood” experience so that they can clearly visualize how AIOps augments their role, rather than eliminating it. Finally, it’s worth ensuring that the solution you’re investing in is future proof. As enterprise goals and initiatives can change year after year, seek out a solution that is flexible, scalable, and agile enough to meet the unforeseen needs of the future.
- The first starting point in developing a successful AI strategy starts with human-based intuition and builds that insight into developing and encoding the right algorithm or right set. To effectively use AI to enhance email marketing, systems need to master the complexities of email technology, manage massive amounts of data, and guide marketers toward sending best practices. We use AI to help solve a customer need. For example, we analyzed a sample of 100,000 email addresses that were invalid due to misspelling, incorrect spacing, or fake accounts. Our team of data scientists began to train AI algorithms to identify these patterns and identify which sets were invalid. We believe the best AI strategy leads with this element of human discovery and then applying AI as a solution and tuning it to fit a need.
- 1) Look at where we build in the cloud, we have access to world-class GPUs. Training eats up a huge amount of compute. Very few large enterprises will invest in that. 2) Conversational is how people will communicate. An AI platform to experiment with models used pre-trained models. Use models for inference. The infrastructure for a data science pipeline those are the middleware angles in terms of software and features. 3) Adaptive intelligent set of applications. AI infused into each vertical application. Within our software AI is being used on the cloud to make the products better for security, ML/DL, products patch themselves. We use ML to analyze all logs of all services to make the cloud better. Pervasive to everything we are doing
- There are 3 key ingredients: correlation, machine learning, and visualization. First of all, to determine where a problem occurs, you need to have a complete, accurate view of your entire environment. That means you need to automatically correlate data from business transactions (like orders, invoices) through application services (such as web servers, databases) and infrastructure (compute, networks, storage). Second, you need to have machine learning analytics that can determine patterns across the entire stack, and finally, you need to be able to visualize the future events. Without a multi-domain, multi-layer and multi-vendor correlation IT teams will have to go through tens of monitoring tools and logs and waste significant time to understand what relates to what. Machine learning based on time-series data will automatically identify the sequence of events that otherwise would be very hard to detect, and visualization will empower IT Operations to take decisive actions to identify and prevent future issues.
Here's who we spoke to:
Opinions expressed by DZone contributors are their own.