Blogs

The advance of technologies today allows businesses to go beyond addressing real-world issues, covering such complicated and hard-to-measure things like customer satisfaction. And because the COVID-19 outbreak has led to the highest rates of technology adoption for e-commerce purposes, it is predicted that by 2025, AI and ML technologies will control about 95% of all customer interactions. 

Because AI and ML are known for helping e-commerce companies with the personalization of services, the use of these technologies leads to increased sales. The growth in sales becomes possible with a product recommendation based on a user's browsing history and buying patterns.  

What's more, AI and ML allow industry players to invent new strategies of communication with customers and keep them engaged. And with the extremely accurate data produced by AI and ML, it becomes possible to enhance the existing strategies, whether the talk is about personalized replies to customer inquiries, tailored alerts on shopping deals, or reminders on the created wish lists. Also, these technologies perfectly align with successful strategies for finding new customers.

Behavior cloning (BC) is, put simply, when you have a bunch of human expert demonstrations and you train your policy to maximize likelihood over the human expert demonstrations. It’s the simplest possible approach under the broader umbrella of Imitation Learning, which also includes more complicated things like Inverse Reinforcement Learning or Generative Adversarial Imitation Learning. Despite its simplicity, it’s a fairly strong baseline. In fact, prompting GPT-3 to act agent-y is essentially also BC, just rather than cloning on a specific task, you're cloning against all of the task demonstration-like data in the training set--but fundamentally, it's a scaled up version of the exact same thing. The problem with BC that leads to miscalibration is that the human demonstrator may know more or less than the model, which would result in the model systematically being over/underconfident for its own knowledge and abilities.

For instance, suppose the human demonstrator is more knowledgeable than the model at common sense: then, the human will ask questions about common sense much less frequently than the model should. However, with BC, the model will ask those questions at the exact same rate as the human, and then because now it has strictly less information than the human, it will have to marginalize over the possible values of the unobserved variables using its prior to be able to imitate the human’s actions. Factoring out the model’s prior over unobserved information, this is equivalent[1] to taking a guess at the remaining relevant info conditioned on all the other info it has (!!!), and then act as confidently as if it had actually observed that info, since that's how a human would act (since the human really observed that information outside of the episode, but our model has no way of knowing that). This is, needless to say, a really bad thing for safety; we want our models to ask us or otherwise seek out information whenever they don't know something, not randomly hallucinate facts.

In theory we could fix this by providing enough information in the context such that the human doesn’t know anything the model doesn’t also know. However, this is very impractical in the real world, because of how much information is implicit in many interactions. The worst thing about this, though, is that it fails silently. Even if we try our best to supply the model with all the information we think it needs, if we forget anything the model won't do anything to let us know; instead, it will silently roll the dice and then pretend nothing ever happened.

The reverse can also happen, where if the model knows more than the human, then it will collect a bunch of unnecessary info which it discards so that its decision is as dumb as the human's. This is generally not as dangerous, though it might still mislead us to the capabilities of the model (we might think it's less knowledgeable than it actually is), and it would use resources suboptimally, so it's still best avoided. Also, the model might do both at the same time in different domains of knowledge, if the human is more knowledgeable in one area but less in another.

Plus, you don’t even need unobserved information in the information-theoretic sense. If your agent has more logical uncertainty than the human, you end up with the exact same problem; for example if it’s significantly better/worse at mental math than the human in an environment where the human/agent can choose to use a calculator provided as part of the environment that costs some small amount of reward to use, even though the agent has access to the exact same info as the human, it will choose to use a calculator too often/not often enough.

This isn’t a purely theoretical problem. This definitely happens in GPT-2/3 currently and is a serious headache for many uses of GPT already--model hallucinations have been a pretty big problem outlined in numerous papers. Further, I expect this problem to scale to even superhuman models, since this BC objective fundamentally does not incentivize calibration. Even as a component of a superhuman agent, it seems really bad if a component of the agent silently adds false assumptions with high confidence into random parts of the agent's thoughts. On the optimistic side, I think this problem is uniquely exhibitable and tractable and the solutions are scalable (superhuman BC would be uncalibrated for the exact same reasons as current BC).

Because an agent that's consistently overconfident/underconfident will get less reward, and reward is maximized when the model is calibrated, the RL objective incentivizes the model to become calibrated. However, RL comes with its own problems too. Making a good reward function that really captures what you care about and is robust against goodharting is hard, and either hand crafting a reward or learning a reward model opens you up to goodharting, which could manifest itself in much more varied and unpredictable ways depending on many details of your setup. A hybrid BC pretrain+RL finetune setup, as is common today (since training from scratch with RL is exorbitantly expensive in many domains) could have the problems of either, both, or neither, depending on the details of how much RL optimization is allowed to happen (i.e by limiting the number of steps of tuning, or having a distance penalty to keep the policy close to the BC model, etc).

I think it would be promising to see whether miscalibration can be fixed without allowing goodharting to happen. In particular, I think some kind of distance penalty that makes it inexpensive for the model to fix calibration issues but very expensive to make other types of changes would possibly allow this. The current standard KL penalty penalizes calibration related changes the exact same as all other changes, so I don’t expect tuning the coefficient on that will be enough, and even if it works it will probably be very sensitive to the penalty coefficient, which is not ideal. Overall, I’m optimistic that some kind of hybrid approach could have the best of both worlds, but just tweaking hyperparameters on the current approach probably won’t be enough.

The use of augmented reality in e-commerce solutions gives consumers an immersive experience, allowing them to have real-time interaction with products while remaining in their own environment. An example of such a solution is the AR Shopping App Designed by Emerline for Driving Shoe Sales. 

AR bridges the gap between physical stores and online shopping experiences. This gap is felt even wider due to the Covid-19 pandemic that restricts retail shops from being open and leaves consumers unable to enter stores and physically handle products. 

So how exactly does AR help e-commerce businesses to survive in this complicated period?

Brilliant Use Cases of AR in E-Commerce

While various researches highlight how augmented reality can not only increase engagement but also have a significant impact on conversions for e-commerce brands, Deloitte has gone one step further by claiming that 40% of shoppers would even pay more for a product if they were able to test it through AR technology. And here are some brilliant examples of how an e-commerce business can benefit from the use of AR today.

L’Oreal is not just one of the leading players in the beauty industry but also one of the Beauty Tech companies in the world. L'oreal has undergone a digital transformation that resulted in the delivery of augmented products and services.

After the company acquired AR try-on technology known as ModiFace, it witnesses acceleration of sales, and highly increased conversion rates. The technology has literally transformed the consumer experience, allowing users to try on virtual makeup and hair color. But things didn't end there. In addition, AR opened up space for skin diagnosis and skin shade assessments.

The adoption of innovation by the company ensured its strong position during the COVID-19 outbreak. The 2020 annual report by L’Oreal states that the beauty company saw a ‘remarkable development’. 

DFS

The leading sofa retailer in the UK with more than a hundred showrooms has launched the largest web-based AR solution in 2020, and after the implementation has reaped outstanding results. By offering their customers an opportunity to try on more than 10 thousand sofas and see how they fit their homes, the retailer easily survived the crisis caused by the pandemic. To be exact, the use of AR and 3D technology allow the company to not just keep but increase the conversion rates up to 112% in December 2020. 

ASOS

ASOS, an e-commerce retailer with over 22.3 million active customers and more than 85000 products on offer, accelerated the use of AR technology during the Covid-19 pandemic to eliminate the need in shooting products in a studio. In collaboration with Zeekit, the company took advantage of AR to simulate real-life photography of more than 500 products on different models each week.

Chatbots

Chatbots allow e-commerce platforms to automate interactions with customers, also adding to more personalized experiences, whether with the help of textual or auditory methods. To be more exact, these are exactly chatbots that allow e-commerce stores to provide 24x7 assistance. 

The important thing to mention here is that these bots not just help with customer queries, but also positively affect buying decisions of users, contributing to the clarity and better navigation on websites with lots of products on offer.  

One more advantage chatbots have on offer is the ability to gather customer input, analyze it, and provide valuable info for better decisions. 

We’ll get back to chatbots later to view this technology and its beneficial impact on business. And now, let’s explore the next brilliant use case of AI and ML in e-commerce.

Hyperlocal ecommerce

We live in a time when hyperlocal commerce is on the rise: it's widely used by e-commerce businesses for the creation of multiple sub-stores under customer geo-location data, in this way allowing to carter the unique needs of every individual customer group. This contributes to a better shopping experience and faster delivery of goods - both highly valued issues. Moreover, hyperlocal commerce serves as a booster to sales by offering bestsellers to users based on the local demand analysis. 

Headless commerce

Used for the delivery of a sound content experience and efficient management of the functionality, headless commerce serves as a good add-in to an e-commerce solution. It can be used by a variety of e-commerce business lines, whether brands that place a focus on the delivery of experiences, companies selling lifestyle products, brands that promote with influencers, etc. - applied as a part of business strategy, it allows creating seamless and personalized shopping experience, unique brand identity, and delivery of customized services.

Localized content

Localized content is a powerful tool in e-commerce that allows businesses to reach wider audiences without any hindrance of location or language. The thing is that multilingual e-commerce solutions allow automatically translating a website into hundreds of languages, just in a few clicks.

There is no doubt that Artificial Intelligence and Machine Learning will continue to assist e-commerce businesses in the areas of website and app optimization, customer experience, personalized services, customer relationship management, warehousing management, etc. Since AI and ML keep evolving, they will continue to benefit the e-commerce industry.

But AI and ML are not the only contributors to the efficiency of e-commerce. So now, let’s see what’s the impact of Augmented Reality on the sector.

D-Series

General purpose compute

The D-series Azure VMs offer a combination of vCPUs, memory, and temporary storage able to meet the requirements associated with most production workloads.

The Dv3 virtual machines are hyper-threaded general-purpose VMs based on the 2.3 GHz Intel® XEON ® E5-2673 v4 (Broadwell) processor. They can achieve 3.5 GHz with Intel Turbo Boost Technology 2.0.

The Dv4 and Ddv4 virtual machines are based on a custom Intel® Xeon® Platinum 8272CL processor, which runs at a base speed of 2.5Ghz and can achieve up to 3.4Ghz all core turbo frequency. The Dd v4 virtual machine sizes feature fast, large local SSD storage (up to 2,400 GiB) and are well suited for applications that benefit from low latency, high-speed local storage. The Dv4 virtual machine sizes do not have any temporary storage.

The Ds, Dds, Das, Dads, Dps, Dpds, Dpls, and Dplds VM series support Azure Premium SSDs and Ultra Disk storage depending on regional availability.

Example workloads include many enterprise-grade applications, e-commerce systems, web front ends, desktop virtualization solutions, customer relationship management applications, entry-level and mid-range databases, application servers, gaming servers, media servers, and more.

E-Series

Optimized for in-memory applications

The E-series Azure VMs are optimized for heavy in-memory applications such as SAP HANA. These VMs are configured with high memory-to-core ratios, which makes them well-suited for memory-intensive enterprise applications, large relational database servers, in-memory analytics workloads etc.

The Ev3-series VMs range from 2 to 64 vCPUs and 16-432 GiB of RAM, respectively.

The Epsv5 and Epdsv5 VM series feature the Ampere Altra 64-bit Multi-Core Arm-based processor operating at up to 3.0GHz frequency. The Ampere Altra processor was engineered for scale-out cloud environments and can deliver efficient performance to reduce overall environmental impact.

The Es, Eds, Eas, Eads, Ebs, Ebds, Eps, and Epds VM series support Azure Premium SSDs and Ultra Disk storage depending on regional availability.

Example workloads include SAP HANA (e.g., E64s v3, E20ds v4, E32ds v4, E48ds v4, E64ds v4), SAP S/4 HANA application layer, SAP NetWeaver application layer, and more broadly memory-intensive enterprise applications, large relational database servers, data warehousing workloads, business intelligence applications, in-memory analytics workloads, and additional business-critical applications, including systems that process transactions of a financial nature.

N-Series

GPU enabled virtual machines

The N-series is a family of Azure Virtual Machines with GPU capabilities. GPUs are ideal for compute and graphics-intensive workloads, helping customers to fuel innovation through scenarios like high-end remote visualization, deep learning, and predictive analytics.

The N-series has three different offerings aimed at specific workloads:

• The NC-series is focused on high-performance computing and machine learning workloads. The latest version—NCsv3—features NVIDIA’s Tesla V100 GPU.

• The NDs-series is focused on training and inference scenarios for deep learning. It uses the NVIDIA Tesla P40 GPUs. The latest version - NDv2 - features the NVIDIA Tesla V100 GPUs.

• The NV-series enables powerful remote visualization workloads and other graphics-intensive applications backed by the NVIDIA Tesla M60 GPU.

NCsv3, NCsv2, NC and NDs VMs offer optional InfiniBand interconnect to enable scale-up performance.

Example workloads include simulation, deep learning, graphics rendering, video editing, gaming and remote visualization.

With estimates of return deliveries reaching $550 billion by 20203, trying to minimize returns will be a primary focus for retailers of all sizes moving forward. Improving customer service has become a lot more metric-driven over the last two decades. The latest customer service stats point to one key initiative: growing customer happiness predictably and at scale.

One way organizations are tackling this challenge is with artificial intelligence.

Companies are launching conversational AI Agents – you might know them as retail chatbots – to help customers throughout the online shopping experience to answer questions a person might have. With an easily accessible shopping assistant, consumers can feel empowered in their decisions and be less likely to have surprises when they receive an item.

AI Agents can take cues from real-time behavior to be able to anticipate when a person is not 100% sure of a decision. For instance, if a person has clicked on the size guide a few times, the company would presume that she is curious about which size she should get.  A virtual agent could preemptively reach out offering advice (This dress is small and will shrink a little in the wash. I’d recommend ordering up a size!). Research has found that sizing issues are a domineering reason that people return items, with 30% of items returned as they were too small and 22% returned for being too big4. If a virtual agent is deployed to help in these situations alone, the impact on the returns volume would be immense. 

Immediate responses to common customer service FAQs

Customer service chatbots are considered to be one of the most widespread use cases for AI, and the e-commerce sector is no exception. They cover a range of issues, from product care and return policies to warranty information and troubleshooting. Providing immediate answers to common questions, e-commerce chatbots significantly save a user's time, which makes this technology highly wanted and beneficial to business. 

Personal shopping and product discovery

AI-powered chatbots define customer preferences and generate personal product recommendations. They also become a number one solution during the holiday shopping season, helping users to find the perfect gift for everyone based on price range, interests, and other criteria. 

Conversational commerce

Chatbots are known for providing a seamless add-to-cart and checkout experience that takes place during a conversation. 

Return prevention

Studies show that over 30% of all online purchases are returned. So, when a user shows behavior that indicates a return is likely to take place, a chatbot can preemptively intervene to prevent a return from ever happening. For example, a user can add two same items in the cart but in different sizes, and a chatbot can intervene to help in choosing the right size. 

Order management

Human agents should not be involved in such mundane tasks as making small changes to an order or tracking the status of delivery because it is costly and often results in wait times, longer resolution times, and increased customer frustration. So chatbots can serve as an assistant to the completion of such tasks.