txt

Vufind (Dba Deepvu)

  • Deep Learning
  • 0 Case Studies
Deep learning as a service for maximizing margins for brands

Apr 12

Building an Expressive & Accurate Interest Graph

Note: This guest post by Moataz Rashad (@moatazr) first appeared in Mashable in Feb 2012 The interest graph has been gaining increased attention over the past few months. With Google transforming itself into a social company, and even pushing personalized search, it’s clear that the curated web is a reality. Furthermore, the interest graph has much more expressive power in the case of the mobile web. Many mobile interest indicators simply don’t exist on the web, for example, checkins, location data, near field communication, etc. Let’s take a close look at what’s involved when building a complete and expressive interest graph platform, and how the resulting graph can effectively optimize ad/reward targeting. Companies developing interest graph platforms have to overcome the following fundamental challenges. 1. Data Collection There are two approaches to collecting interest data: explicit and implicit. Companies such as Hunch, Blippy and Pinterest have all attempted to address this by asking users to share interests explicitly, with the incentive that users will get more accurate recommendations in return. However, this approach has some psychological hurdles. Why would I tell an engine that I like BMWs, only for it to suggest BMW deals to me soon after? And if I mention that I like sports cars, in general, would it suggest cars that are too pricey or too big for my taste? In order to get accurate recommendations, I should be more specific: I like sports cars, but mostly European ones, and only those that are under $60,000. In other words, the amount of data that I have to volunteer in order to get a decent recommendation or deal is so high that I might as well just search for the product directly. Amazon does a great job at implicit data collection. The site has all the elements of the equation: your purchase history, product search history and even product correlations (people who bought this also bought that). Any app where users have to take the time to populate their own interests will invariably have inaccurate or artificial interest profiles. Facebook encountered this issue with its interest hubs: Most people didn’t take the time to populate their profiles with accurate interests. Even if they did spend the time, they populated the subset of interests that they wanted to project publicly. Google+ Sparks faced the same issues. Interests have to be inferred from normal app usage, where users have opted in. And the app’s normal usage has to provide strong interest indicators. General social networks such as Facebook and Google+ have the luxury of collecting vast data of various types. Specialized social networks such as Foursquare and Pinterest collect data that is heavily biased towards one signal (checkins in the former, and liking photos that belong to certain interest categories in the latter). 2. Noise Filtration Every action online is considered a signal, and almost every signal in the digital world has its fair share of noise, though the noise levels and types vary greatly. For example, comments are extremely noisy (LOL, OMG, etc.), as are Likes/+1s when applied to photos or comments etc. However, Likes/+1 of brand pages, for example, are very reliable interest indicators. Essentially, “noise” becomes valuable, depending on the amount of effort a person puts in. But the degree of correlation varies depending on the person’s behavioral profile. If you take the time to upload a video of yourself skiing, that’s a strong signal. If you simply Like/+1 someone’s skiing photo, it may be that you like the person, or you like skiing, or you are simply trying to get the attention of the poster in order to start a conversation. Repeated checkins at restaurants/bars are strong interest indicators as well. And clearly, reward/deal redemptions and purchases are very strong signals. Machine learning algorithms are typically used to detect noise patterns and spammy comments, etc. Similarly, signal strengths have to be analyzed. To give simple examples, uploading sailboat photos every week is a different level of interest than liking a friend’s sailing photo once in a while. Similarly, checking in at the same sushi restaurant six times last month sends a very different signal than checking in once every two months. However, here’s where the complexity arises. One checkin per month may actually be a strong indicator if the person travels frequently. In other words, the signal strengths calibration algorithm has to be customized according to the person’s behavioral patterns and lifestyle. The beauty is that you don’t have to get it right the first time, since, like any neural network, the engines improve with usage over time. 3. Building the Interest Graph Even after noise filtration engines have been well-trained and continuously re-calibrated to build an interest profile for a given user, constructing an interest graph for a set of users is still challenging. Essentially, the complexity is in aggregating all the signals to form a coherent and reasonably consistent profile. 4. Platform APIs In my mind, this is the biggest challenge, and so far, no company has managed to deliver a solution. Clearly, a lot of companies build their own interest graphs for their own user bases. However, being an interest graph platform means publishing APIs that any app/game can use to personalize ads or commerce to their users. This means that your APIs have to be at such fine granularity that other apps/games can integrate them seamlessly, with no detrimental impact on the user experience. 5. Distribution: Attracting the Apps to Use the APIsAs in any B2B sales cycle, the first few customers are the hardest to acquire. In this case, they are also the most important, since their impact on engine accuracy is very significant. It’s very clear that interest graphs are at the core of the curated mobile web, and will be a key driver for mobile commerce for years to come. Many apps are already building relatively accurate interest graphs of their user bases. Ultimately, the companies that build scalable interest graph platforms with APIs that map to numerous apps and games will dominate the mobile commerce ecosystem.

Read full article >


Sep 12

Nintendo’s scarcity, and ensuring on-shelf availability

By Moataz RashadRecently Nintendo has been facing a problem many consumer electronics manufacturers envy– consumers love their game consoles! But now the iconic company has a brand image problem, ie. why can’t it accurately forecast the demand, and ensure its new products are available at its retail partners.  See this and this.In a nutshell, there are two problems in play here: a) Demand forecasting, and b) On Shelf Availability. In this case, Nintendo is struggling with both. Firstly, they under-forecasted initial demand for the new product at the launch date that they selected. Secondly, subsequent to the actual launch, they didn’t coordinate their supply chain with the retailers’s data (point of sale and distribution network) to ensure On Shelf Availability for the product so its eager customers can buy it.On Shelf Availability is a sub problem of demand forecasting that deals with optimizing the likelihood that a customer who walks into a store (physical or online) searching for a product X, will find that product both in-stock and, in case of physical stores, actually on the shelf to complete the purchase.It is clearly the one metric that ensures no sales are lost. If a customer intends to buy, they will find the product available at the time and location of their choice.  This problem is a key supply chain problem that’s been around for decades. It requires close real-time collaboration between the manufacturers and the retail channels with daily updates of on-hand inventory and point-of-sale transactions. But most importantly, it requires deep-learning to solve it at scale. Unfortunately, traditional machine learning techniques deliver only 55–65% accuracy which costs millions in lost sales. They are unable to ingest the numerous external signals that impact demand, and they are unable to handle the massive volume of data the tier1 manufacturers and tier1 retailers deal with. So they’re left with building numerous disparate models instead of one large monolithic deep learning engine that can learn from all the patterns and transfer learnings from one store’s patterns to a sister store etc.A deep learning solution is able to handle all kinds of data feeds that impact the customers’ demand patterns, including economic and demographic data which vary by vertical and product category. It also ingests the PoS and inventory data along with daily feedback from the manufacturer and retailer in order to adapt and self-tune its predictions in real-time. That’s what Nintendo and comparable major enterprise manufacturers need in order to address these key supply chain challenges that could potentially make or break the company.

Read full article >

Sep 12

Future of commerce: AI replaces BI

For the past couple of years, many well-respected folks have opined on the future of retail, claiming that all brick and mortar retail is dead, dying, or in a vegetative state, and that the future of shopping therefore lies entirely online! See “future of shopping” and “retail stores will die”I certainly wouldn’t mind that notion materializing at all, since for obvious reasons, the more commerce moves online–which is an irreversible trend–the better for us at Vufind, a startup building the AI layer for eCommerce.However, I think that the hypothesis is flawed. The future of retail is AI-commerce, i.e. smarter shopping experiences in every sense of the word, whether it happens online, offline, or hybrid. “Smart” as opposed to “not so smart” commerce, as in what you often see now i.e : a) having to wait in line to buy a staple product for which there is little choice, minimal price variation, and there’s no emotional investment in the purchase b) being chased after you’ve left a store (offline or online) and abandoned an item to be shown the same exact item over and over (aka retargeting) or c) the seller keeps telling you those who bought this also bought that, as if all shoppers are automatons (aka collaborative filtering)See this emarketer presentation for more detailed proof points on consumer data regarding how they view online vs. offline shopping.It is indisputable that online retailers have much more attractive cost structure, and are advantaged by the digital consumer and product data that can drive decisions in real-time at nano-second speeds. However, offline retail has some advantages as well, including immediate delivery, try before you buy, impulsive buys, and lower return rates.Some online-offline hybrid experiences are already happening. For an example of online companies expanding offline, just monitor the investment moves of Amazon, one of the most forward-thinking online retailers globally, and you’ll observe that they are serious about online-offline interplay. Similarly, Apple retail has a hybrid show-room/store model which is being copied by other consumer electronics companies. On the other hand, a forward-thinking brick and mortar retailer such as Nordstrom is experimenting with exciting online-offline hybrid experiences with the smart mirror for example. It will be intriguing to see the A/B test results of these efforts as they will pave the way for even more fascinating experiences. Courtesy: Nordstrom-ebay fitting room So what is smarter commerce, or AI commerce? AI commerce has one broad goal: matching the shopper with their desired product by factoring in their style, brand preference, price range, size, fit, and delivery time, in order to effect a frictionless speedy transaction.If a brick and mortar retailer builds a sophisticated site/app but they are primarily interested in driving traffic to their physical stores, then they aren’t fostering an AI commerce experience. By the same token, if an online-only retailer gets user feedback in favor of same-day delivery and ignores this input, then they aren’t developing an AI commerce experience. So what would an AI commerce experience entail?Let’s start with some definitions since AI is such a land-mine these days, what with everyone from famous loved entrepreneurs such as Elon Musk and Bill Gates to the occasional scientific journalist speaking on the topic.AI in this context refers to the field of Artificial Intelligence; software algorithms that use pattern recognition, computer vision, deep-learning neural networks, or whatever other techniques we practitioners come up with, to produce intelligent decisions that impact the shopper’s user experience. Such decisions include recommendations, merchandizing, dynamic pricing, promotions, placement of products on sites/apps/store-window, or whatever decision the marketing/product management executive needs the algorithm to assist with that day. In other words, we’re not talking about an artificially intelligent entity such as a robot, vehicle, drone, or anything else that can do any harm to people in the real world. And we’re certainly not talking about “Strong AI” which is where most of the fears of the critics revolve, where the algorithms would decided to learn from data outside their domain, and in fact will consume all accessible online knowledge, and for some reason they’ll take special interest in learning about destroying humanity!In AI-commerce, we’re only discussing deep-learning software running in the cloud with no tentacles in the physical world, and where it’s trained on where to look, what to look at, and what type of output we’d want it to produce, and it’s only domain is commerce. This begs the question– how is AI different form Business Intelligence (BI) which has been the traditional software layer of eCommerce? At the highest level, the fundamental difference is that BI is static or doesn’t learn from the data, while AI’s missions is to continuously improve and learn from the data to get “smarter”. In other words, Business Inteligence is a misnomer, it’s really business reporting, or business stats– dry static stale reports on what’s already happened a while ago.So what do these AI algorithms work off of? the online shopping experiences have the upper hand here, because every action is already digital, cloud stored, and hence trackable and learnable in real-time. Every action including product impression, image click, recommendation click, purchase, like, save, share, cart-abandonment,email open, email link click, etc, is fed to the algorithms with the goal of making the shopper’s experience smarter and more pleasing.We believe a true AI-commerce experience should accommodate most of the following:1. Smarter visual recommendations: A smarter experience means that you get to see a recommendation that visually matches your style preference so much that it persuades you to purchase the recommended item rather than the item you originally searched for. We get super stoked when we count those transactions.2. Smarter personalized recommendations: If someone is a loyal repeat user of an etailers site, then they do in fact expect a personalized experience that accounts for more than gender and age group, regardless of whether or not they use a social login. If they have purchased before, browsed, liked, and abandoned carts before, they expect the site to be smart enough to know their taste preferences and which products they would like to see. How is this different than (1) ? A user might be looking at a heather grey lambswool sweater, and you could show her ski goggles and cycling shoes, because you’ve learned that she’s into the sports associated with these items and haven’t purchased them in over 9 months. You might even show her girl’s soccer shoes because she bought a pair for her daughter last February. Or you might surprise her with an item that she’s never bought from your site before because it matches her style.3. Smarter merchandizing: the more AI-powered decisions the retailer employs, the smarter the product mix in their catalog. As such, their buyers would know exactly what to stock each season based on transaction trend analytics, thus eliminating waste across the supply chain therefore saving time, and resources.4. Psychic Personal Shopper Apps (PSA): A truly intelligent PSA wouldn’t just have a warm voice and uses Google and Amazon to run searches for you and send you a list, it would know what you’d want and would just get it delivered to your door step. And if you dislike the product, you tap a button and it’ll get it returned for you as well. It’ll know that it’s time to re-order socks, or that winter is coming 2 weeks earlier, so it’ll order those thermal under-shirts. And it’ll buy a different tasteful gift for your best friends’ birthdays each year, and it’ll ask you what the friend said about it so it gets better at gifting. Most importantly, it’ll do all that algorithmically without any *crowd-sourcing* or asking you to fill out forms.5. Smarter “show-rooms”: Recall when show-rooming was considered a threat, well it isn’t, it’s part of the shopping experience now, and every online and offline retailer designs for it. However we think in the very near future, the physical shops will evolve into AI powered show-rooms, amazingly chic physical spaces that allow shoppers to touch, feel, try on, learn about the origin of the product, or quickly browse similar products to explore. But not necessarily wait in line or carry it home, unless in fact they do need it that exact moment.6. Smarter payment: Some of the amazing international etailers we work with offer cash payment upon delivery. Some have added BitCoin. Some experimenting with pick up at partner physical stores. Retailers should certainly make it as frictionless as possible to pay for the item you’ve already invested the time and emotional energy to shop for and decided to purchase. So smarter payments means choice, speed, convenience, and quite frankly none of those hideous 2.9% fees, which I personally think should be eliminated.7. Social: Many interesting developments there, but that’s for another article.The history of commerce goes back thousands of years. It’s catering to a human need that’s existed since the oldest civilizations used the barter system. The future of commerce, however, is truly awesome; it’s intelligent commerce powered by AI, easier, faster, cheaper, more relevant and more delightful. It’s already happening.Are you excited about smarter commerce?–would love to see your thoughts below.

Read full article >

Aug 07

Will AI “save” manufacturing?

By Scott Lyon“Manufacturing Loses Most Jobs Since Election” reads the latest headline.According to the latest job report, the U.S. manufacturing sector once again reported job declines (4,000 jobs lost).  This further extends an unfortunate erosion in our country’s manufacturing base and in fact Institute of Supply Management (ISM) surveys suggest job losses will continue.Although efforts by the current administration to reverse the trend toward “off-shoring” as well as aggressive incentives prepared by the states (e.g. check out Wisconsin’s just-announced Foxconn package) may prove helpful in the short-term, however, clearly the manufacturing sector needs to reinvent itself and develop new market opportunities.One increasingly popular theme for doing so involves Artificial Intelligence (AI).  Namely, could deeper machine learning enable manufacturing companies to either reduce costs and/or drive incremental revenues without launching new products or securing additional customers? There are numerous conceptual opportunities suggesting the answer is yes, here are three recent compelling use cases:Two ex-Apple engineers have created hardware which takes photos at critical junctures of a manufacturing line and can pinpoint in real-time when devices are being assembled incorrectly (all without requiring on-site visits by managers or quality control inspectors);GE launched an internal IT initiative which is now providing advanced analytics to external customers for example advanced sensors for their oil / gas clients can better predict blade health and determine optimal replacement times;One of the leading semiconductor companies is leveraging DeepVu’s AI engine  to determine the “optimal production allocation” for maximizing chip revenue and optimizing yield amongst multiple contract manufacturers;In each of these three examples, manufacturers are leveraging their existing data and service offerings to drive either reduced equipment failures or greater revenue-per-channel.  

Read full article >

POSTS FROM RELATED COMPANIES
Microsoft - Apr 08 - Blog Post:
How international health care organizations are using bots to help fight COVI...

Health care organizations are using Microsoft’s Healthcare Bot service to help screen people for potential coronavirus infection and treatment.

Microsoft - Apr 08 - Blog Post:
How international health care organizations are using bots to help fight COVI...

Health care organizations are using Microsoft’s Healthcare Bot service to help screen people for potential coronavirus infection and treatment.

A Us Based Big Data Intelligence Company - Apr 08 - Blog Post:
COVID-19 Use Cases, Goals, Challenges, Impact

A breakdown of COVID-19s impact on healthcare and how AI is able to assist in improving efficiency, deterring unneeded ER...

A Us Based Big Data Intelligence Company - Apr 08 - Blog Post:
COVID-19 Use Cases, Goals, Challenges, Impact

A breakdown of COVID-19s impact on healthcare and how AI is able to assist in improving efficiency, deterring unneeded ER...

NVIDIA AI - Apr 08 - Blog Post:
Seeing AI to AI: NVIDIA Deepens Ties with Top Research Center

NVIDIA has become a shareholder in DFKI, the largest AI research center in Germany.

Keep updated on all things AI